IT202100001424A1 - METHOD AND SYSTEM FOR PREDICTING THE FAILURE OF A MONITORED ENTITY - Google Patents

Description

METHOD AND SYSTEM FOR PREDICTING THE FAILURE OF AN?ENTIT?

MONITORED

DESCRIPTION

TECHNICAL FIELD

The present invention relates to the sector of computer systems. In detail, the present invention relates to a method and a system for predicting the failure of an entity monitored.

STATE OF ART

In various sectors of activity? are implemented systems of great complexity?, each divided into a plurality? of units in turn divided into a plurality? of components. Examples of such complex systems include industrial manufacturing plants, public transportation systems, telecommunications networks, computer networks, data centers, etc.

The complexity? of these systems makes it extremely difficult and laborious to control and verify the status of the various units? and components that make up each system. Consequently, periodic maintenance operations are often implemented regardless of the actual need? of the same and/or the systems are implemented with a considerable redundancy to avoid that the malfunction of a critical component compromises the entire system. As ? evident such solutions, only partially limit the disadvantages mentioned and lead to a substantial increase in the cost of production and operation of these systems, so? how the complexity? of the same.

In an attempt to solve these problems, methods based on an automatic analysis of complex systems, in particular based on artificial intelligence, have been proposed in the art in order to predict a possible malfunction of the same.

For example, WO 2019/125445 and US 6,125,105 propose methods based on an artificial intelligence intended to predict malfunctions in a network of computer systems, in particular as regards the exchange of data between elements of the network.

US 2013/132001 and US 2012/025997 instead propose methods based on artificial intelligence intended to predict the malfunction of mechanical or electromechanical devices.

Furthermore, methods based on artificial intelligence are known for predicting malfunctioning situations of complex systems not only influenced by mechanical and/or electrical factors, but also economic/social factors. For example, JP 2020/095583 describes an artificial intelligence-based method intended to provide a prediction of the probability of bankruptcy of a company.

The Applicant has observed that the proposed methods, although based on robust artificial intelligence algorithms, are strongly influenced by the composition of the data used during the training of the artificial intelligence algorithms.

In particular, the Applicant has found that the known training data processing techniques do not allow to define an optimal training data set, especially in the case of large non-uniform data sets, with a consequent negative effect on the training quality. of the operating predictions generated by known systems.

OBJECTS AND SUMMARY OF THE INVENTION

? object of the present invention is to overcome the drawbacks of the prior art.

In particular, ? object of the present invention to provide a method and a related system for predicting an optimized malfunction.

A further object of the present invention ? that of proposing a method and system for predicting a malfunction capable of providing reliable malfunction predictions on a plurality of? of periods of future times.

A further object of the present invention ? that of proposing a method and system for predicting a malfunction capable of automatically creating a set of optimal training data starting from a set of available data.

These and other objects of the present invention are achieved by means of a system incorporating the characteristics of the attached claims, which form an integral part of the present description.

According to a first aspect, the present invention ? directed to a method to predict a malfunction of an entity. In particular, the entity? ? characterized by a plurality of parameters.

Does the method expect to predict a malfunction of the entity? by means of an artificial intelligence system which is provided with data relating to the parameters of the entity, in which the artificial intelligence system is trained using training data.

Advantageously, the training data is selected according to the following steps:

to. acquire a set of training data, wherein the set of training data comprises a plurality of of entity data groups, each entity data group? including values referring to characteristic parameters of a respective entity? of training;

b. select entity data? included in entity data groups? of a first subset of entities? of training of the set of training data; c. train the artificial intelligence system using entity data? of the first subset of entities? of selected training;

d. determine a probability distribution? of malfunction of the entities? of training of the set of training data, running the artificial intelligence system trained with the data of entity? of the first subset of entities? of training, e

And. determine the training data by selecting data entity? of entity? training belonging to a second subset of entity? of training of the set of training data in such a way that the second subset of entities? training present a distribution of probabilities? of malfunction equal to that calculated in point d.

In the present the expression ?entity? of training data set? indicates one (or more, depending on the case) entity? of training of which at least a respective group of data of entity? of training is included in the set of training data.

Thanks to this solution ? It is possible to obtain particularly accurate malfunction predictions even when the training data set used for artificial intelligence training is ? highly unbalanced. In particular, ? the same artificial intelligence system that creates the optimal training data completely automatically.

In one embodiment, the method further comprises the step of: - determining a first group of entities? of training of the set of data of slack that have suffered a malfunction and a second group of entity? of the training data set that have not suffered a malfunction.

When the second group ? much greater than the second group the step of determining a distribution of probabilities? of malfunction of the entities? of training of the set of training data expects to determine the distribution of probabilities? of malfunction of the entities? of training of the second group.

Also, the step of determining training data by selecting entity data? of a second subset of entities? of the training dataset includes:

- select a third subset of entities? training of the second group in such a way that the third subset of entities? training present a distribution of probabilities? of malfunction equal to that calculated in points d, e

- define the second subset of entities? of training of the set of training data as a union of the entities? of training of the first group and of the third subset.

In the present, the expression ?much greater? indicates a ratio equal to or greater than 1:5.

In this way ? You can obtain balanced training data from a heavily biased training dataset quickly and efficiently.

In one embodiment, the step of determining a probability distribution is of malfunction of the entities? of training of the training set includes:

the. define a plurality? of probability threshold values;

ii. determine at least one of a number and a list of entities? of training of? set of data of training whose probability? of malfunction? less than or equal to a selected threshold value, running the AI system,

iii. to select a new threshold value of probability? of said plurality? of probability threshold values?,

iv. iterate steps i - iii, until all probability threshold values have been selected, e

v. calculate the probability distribution? of malfunction of the entities? training set based on the results provided at each iteration of step ii.

Preferably, the steps of include the steps of:

ii_a. select entity data? of an intermediate subset of entities? training, said intermediate subset of entity? training including entity? included in the list of entities? of training determined in step ii and entity? of training set that have suffered a malfunction;

ii_b. train the artificial intelligence system using entity data? of the intermediate subset of entities?,

before performing steps iii - v.

In one embodiment, the step of determining a probability distribution is of malfunction of the entities? of training of the training set includes:

- calculate a probability? of malfunction for each entity? of training of the training data set, running the artificial intelligence system, e

- combine the odds? malfunction calculated to determine the distribution of probabilities? of malfunction of the entities? of training data set.

In one embodiment, the step of determining a probability distribution is of malfunction provides for:

- calculate a plurality? of probability distributions? of malfunction of the entities? of training of? set of training, where each distribution of the probabilities? of malfunction of the entities? of training of? training set ? calculated by means of groups of data of the entities? of training referred to the same time interval.

In addition, the step of determining training data involves:

- select entity data? of a second subset of entities? of training of the set of training data in such a way that the second subset of entities? of training present a plurality? of probability distributions? of malfunction equal to the plurality? of probability distributions? of malfunction calculated for the entities? of training data set.

Has the Applicant determined that the methodologies proposed above make it possible to effectively determine the probability distribution? failure of the whole in a particularly effective way.

In one embodiment, the step of selecting entity data? of a first subset of entities? of the training dataset includes:

- determine a first group of entities? data set training data that has malfunctioned;

- randomly sample a second group of entities? training group of the set of training data that have not suffered a malfunction, where the second group includes a number of entities? training between 1 and 5 times a number of entities? of training of the first group, and - compose the first subset of entity? of training of the set of training data as a union of the entities? of training of the first group and of the second group.

In this way, ? It is possible to obtain an initial reduction of an unbalanced training data set that allows the artificial intelligence system to be trained in a functional way.

In one embodiment, the step of predicting a malfunction of the entity? through an artificial intelligence system plans to use an automatic learning algorithm, selected from:

-Random Forest,

- XGBoost;

- Logistic Regression, e

- a neural network.

The Applicant has determined that these machine learning algorithms, in particular the Random Forest algorithm, allow for optimal performance. For example, the Random Forest algorithm allows to obtain results characterized by an AUC factor (acronym of Area Under Curve - area under curve ROC, acronym of Receiver Operating Characteristic ? operating characteristic of receiver) between 88% - 93%.

In one embodiment, the step of predicting a malfunction of the entity? through an artificial intelligence system expects to calculate a plurality? of probability of malfunction of the entity?, each probability? of malfunction of? Entity? referring to a respective future time interval.

The malfunction forecasts provided by the system are highly reliable even over very long future time periods, for example between one and five years.

A different aspect of the present invention relates to a system for predicting a malfunction of an entity. In particular, the entity? ? characterized by a plurality of parameters, the system including:

- an interface module configured to receive data relating to said entity? and providing a malfunction prediction result;

- a data processing module configured to acquire and process data relating to entities? of training and the data relating to this entity? supplied to the interface module, e

- an artificial intelligence module configured to run an artificial intelligence algorithm that uses the data processed by the data processing module to predict a probability? malfunction of the entity.

Advantageously, the modules of the system are configured to implement the method according to any one of the preceding claims.

Further characteristics and purposes of the present invention will become clearer from the following description.

BRIEF DESCRIPTION OF THE DRAWINGS

The invention will come described hereinafter with reference to some examples, provided for explanatory and non-limiting purposes, and illustrated in the attached drawings. These drawings illustrate different aspects and embodiments of the present invention and, where appropriate, reference numerals illustrating similar structures, components, materials and/or elements in different figures are indicated by like reference numerals.

Figure 1 ? is a block diagram of a malfunction prediction system according to an embodiment of the present invention arranged to provide one or more predictions of malfunction of a? entity? monitored;

Figure 2 ? a flowchart of a training procedure implemented by the malfunction prediction system of Figure 1 ;

Figure 3 ? a qualitative graph of a probability distribution? of malfunction calculated by the malfunction prediction system of Figure 1, e

Figure 4 ? a flowchart of a malfunction prediction procedure implemented by the malfunction prediction system of Figure 1.

DETAILED DESCRIPTION OF THE INVENTION

While the invention? susceptible to various modifications and alternative constructions, some preferred embodiments are shown in the drawings and will be described in detail below. It must be understood, however, that there is no no intention of limiting the invention to the specific embodiment illustrated, but, on the contrary, the invention is intended to cover all modifications, alternative constructions, and equivalents which fall within the scope of the invention as defined in the claims.

The use of ?for example?, ?etc.?, ?or? indicates non-exclusive alternatives without limitation unless otherwise indicated. Using ?include? means ?includes, but not limited to? unless otherwise indicated.

With reference to Figure 1, a malfunction prediction system based on artificial intelligence according to an embodiment of the present invention ? referred to as ?system 1? in the following, for brevity? ? comprises an interface module 11, a data processing module 12 and an artificial intelligence module 13, connected to each other to exchange data? for example, in binary format.

The interface module 11 ? connected to one or more user devices 2 (for example, computers or IT systems) in order to exchange data with them. Furthermore, the interface module 11 ? connected to one or more sensors 3. In particular, each sensor 3 ? configured to provide a measure of a physical quantity associated with the operation of an? entity? monitored 4.

For this purpose, the interface module 11 includes necessary hardware resources, such as one or more processors, volatile memories, non-volatile memories, modems, etc.

For example, the entity? monitored 4 ? an industrial plant comprising a plurality? of automated processing stations. In this case, the sensors 3 can comprise one or more? between temperature sensors, pressure sensors, weight sensors, cameras, thermal imaging cameras, voltage and/or current sensors, flow sensors, etc. according to the type of industrial plant and/or the processing stations included in the same. Additionally, the sensors 3 may include processing devices such as PLCs, microcomputers, microcontrollers, etc. In this case, are these processing devices configured to supply operating information of the industrial plant and/or of the processing station thereof to which they are connected? for example, production volume, uptime? of the components included in the processing stations, etc.

The interface module 11 ? configured to supply the data received from the user devices 2 and from the sensors 3 to the data processing module 12. Furthermore, the interface module 11 ? connected to the artificial intelligence module 13 to exchange data with it. Advantageously, the interface module 11 ? configured to transmit to the artificial intelligence module 13 prediction requests received from a user device 2 and/or input data to be used for predictions as described below. In addition, the interface module 13 ? configured to receive output data generated by the artificial intelligence module 12 to be forwarded to one or more? corresponding user devices 2.

The data processing module 12 ? connected to one or more databases 5, or more? in general data sources, in order to exchange data with them. In particular, the databases 5 include a plurality? of entity data? db. Each entity data? DB includes values of a plurality? of characteristic parameters of a respective entity? (not illustrated).

The data processing module 12 ? configured to receive, combine and process the data acquired by the databases 5 and by the interface module 11, so that they can be used adequately by the artificial intelligence module 13 in order to provide a malfunction prediction. For this purpose, the data processing module 12 comprises necessary hardware resources, such as one or more processors (CPU), graphics processors (GPU), signal processors (DSP), volatile memories, non-volatile memories, modems, etc.

Artificial intelligence module 13 ? configured to elaborate data processed by the processor module 12 and/or supplied by the interface module 11 in order to predict a probability? of malfunction of? Entity? monitored 4. The artificial intelligence module 13 ? configured to run an AI algorithm, preferably, a machine learning algorithm ? or Machine Learning (ML). In a preferred embodiment of the present invention, the artificial intelligence module 13 is configured to run the machine learning algorithm, preferably a random forest algorithm ? or Random Forest.

For this purpose, the artificial intelligence module 13 comprises necessary hardware resources, such as one or more processors (CPU), graphics processors (GPU), signal processors (DSP), volatile memories, non-volatile memories, modems, etc.

The system 1 just described ? configured to run a method to predict a malfunction of an? entity? according to an embodiment of the present invention.

In particular, the method comprises a learning procedure 100 and a prediction procedure 200.

The learning procedure 100 has the purpose of training the artificial intelligence algorithm AI in order to perform reliable malfunction predictions.

In detail, the learning procedure 100 envisages acquiring a set of entity data? Entity-Referred DBs? of training from the databases 5 (block 101). each entity training activity has characteristics and purpose substantially corresponding to, or at least comparable to, the characteristics and purpose of the entity? monitored 4. For example, the entities? of training are industrial plants that produce the same or similar finished or semi-finished product to the finished or semi-finished product from the industrial plant corresponding to the entity? monitored 4.

Advantageously, the entity data? of training acquired refer to both entity? of training that have suffered a malfunction is entity? training devices that have not suffered a malfunction. For each entity? of training ? acquired at least one group of entity data. Preferably, for each entity? are groups of entity data captured? refer to two or more predetermined time intervals ? for example, five sets of entity data? referred to? operation of? entity? of training during five consecutive years, ten data referring to as many periods of exercise of the entity? of training separated by a period of rest, etc.

Advantageously, each group of entity data? does it include values referring to characteristic parameters of the respective entity? of training, both tempo variants and tempo invariants. For example, time-independent parameters include structural parameters (dimensions, weight, list of stations and/or components, etc.) and nominal parameters (maximum absorbed/generated power, maximum frequency of operation, maximum load, MTBF of one or several components of the plant and/or of the stations of the plant, etc.). Conversely, do time-dependent parameters include entity operating parameters? of training (instantaneous absorbed power, revolutions per minute of a mechanical/electromechanical component, quantity of input raw material, quantity of finished or semi-finished product on output, etc.). In particular, the time-dependent parameters can include a plurality of of values acquired in different instants of time, a trend of the value of the parameters in said time interval, a variation of the value of the parameters between two or more? instants of time and/or similar information.

Entity data? Acquired DBs are made uniform with each other (block 102). In particular, the entity data? are processed so that each group of entity data? does it include values referring to the same characteristic parameters of the entity? of training.

For example, ? performed a missing value imputation operation as described in Barnard, J., Meng, X.-L., 1999: ?Applications of multiple imputation in medical studies: from aids to nhanes? Statistical methods in medical research 8 (1), 17?36, and a standardization of the data set. Preferably, any missing values from an entity data? ? place at zero and ? performed a standard scaling operation.

Entity data? Standardized DBs are then analyzed and separated into main data, or core data, and secondary data, or non-core data (block 103). Preferably, the master data comprises sets of entity data? having values referring to the same parameters, so as to allow an adequate comparison between the corresponding entities. Otherwise, does secondary data collect data referring to parameters not contained in all entity data groups? of the entity data set.

The secondary data is processed to generate perturbation data (block 104). In particular, perturbation data represents an effect of secondary data on primary data.

The main data is then divided into a set of training data DT and a set of validation data DV (block 105). The master data is split so that the DT training data set is substantially larger than the DV validation data set. Preferably, the DT training data set comprises 80% of the entity data? DB and, therefore, the set of DV validation data includes the remaining 20% of the entity data? db.

The set of training data DT ? then subjected to a process of reduction of characteristics ? or Feature Reductions? (block 106). For example, parameters are removed for which the presence or absence does not change the accuracy of the results by a value greater than 1% when performing a classification operation.

Preferably, the feature reduction process ? an iterative feature removal process ? o iterative feature removal process.

Subsequently, ? selected a subset of entities? of training associated with corresponding groups of data of entity? of the training data set DT (block 107). For example, ? performed a resizing (downsizing) of? set of training data, by means of a random sampling of the entities? of training of the set of training data DT, in order to obtain a subset of entity? of training substantially reduced compared to the totality? of entity? of training data set DT.

In the present the expression ?entity? training data set DT? indicates an?entity? of training of which at least a respective group of data of entity? of training is included in the set of training data DT.

Preferably, in the case of a highly unbalanced training dataset, ? expected to select a subset of entities separately? of training of the set of training data DT that have not suffered a malfunction and a subset of the entities? of training of the set of training data DT that have undergone a malfunction, in particular with different dimensions, and then merged into a single subset of entities? of training. Preferably, the subset of entities? of reduced training? characterized by the proportion between the entities? of training that have suffered a malfunction and the entities? of training that have not suffered a malfunction ? between 1:5 and 1:1, preferably equal to or less than 1:3.

The expression ?highly biased training data set? indicates a training dataset in which the entities? of training that have suffered a malfunction are 20% or less of the entity? training devices that have not suffered a malfunction.

The training data associated with the subset of entities? training data are used as the first subset of training data DT1 (block 108) and ? used to train the artificial intelligence algorithm AI (block 109).

What is the AI artificial intelligence algorithm? trained ? evaluated (decision block 110). In particular, the DV validation dataset and the perturbation data are input to the AI artificial intelligence algorithm and then the malfunction prediction matching is verified for each entity? of training associated with the DV validation data. Preferably, ? verified that the AI algorithm malfunction predictions are not affected by overfitting.

If the validation detects the production of sub-optimal results (output branch N of block 110), the hyperparameters (hyperparameters) of the artificial intelligence algorithm are possibly adjusted so as to optimize the precision and recovery of the malfunction predictions provided by the artificial intelligence algorithm (block 111) and ? repeated the validation step to verify that the AI algorithm works as desired. Preferably, the adjustment of the hyperparameters ? performed by means of a regulation approach based on grid search (Gridsearch) - for example, in a similar way to what described in Bergstra, J., Bengio, Y., 2012: ?Random search for hyper-parameter optimization?, Journal of Machine Learning Research 13, 281?305.

Once the validation detects the production of acceptable results (output branch Y of block 110), ? planned to use the AI artificial intelligence algorithm to calculate a probability? of malfunction of the entities? training data set DT (block 112).

For example, the artificial intelligence algorithm AI ? performed more times. At each iteration the algorithm ? configured to output a number and/or a list of entities? of training of? data set DT whose probability? of malfunction? less than or equal to a respective threshold value.

Preferably, the artificial intelligence algorithm AI ? performed to determine the entity? training with a probability? malfunction equal to or less than 20%, 30%, 40%, 50% and 60%.

even more preferably, the operation of calculating the probability? of malfunction of the entities? training data set training DT ? performed for each time interval for which there are corresponding groups of entity data? of training.

By combining the results of the iterations of the artificial intelligence algorithm AI, a probability distribution is obtained. of malfunction of the entities? of training data DT (block 113) (illustrated qualitatively in Figure 3). Preferably, ? Is a probability distribution defined? of malfunction of the entities? data of training data DT for each time interval for which there are corresponding groups of data of the entities? of training.

Based on the probability distribution? of malfunction cos? defined, ? selected a second subset of training data DT2 starting from training data DT (block 114). In particular, the second subset of training data SDT2 includes data referring to a subset of entities of training of the entities? of training of the set of training DT selected so that the subset of entities? training has a probability distribution? malfunction corresponding to the probability distribution? of malfunction of the totality? of the entities? of training of the training set. Preferably, the second subset of training data SDT2 ? defined so that there is a correspondence with each of the distributions of the probability? of malfunctions calculated for a respective time interval.

The artificial intelligence algorithm AI ? subjected to a new workout based on the second subset of workout data DT2 (block 115).

At the end of the training, the artificial intelligence algorithm AI ? again subjected to evaluation (decision block 116) and adjustment of the hyperparameters (block 117) in case the production of sub-optimal results is detected (output branch N of block 116), in a similar way to what was described above.

At the end of the training procedure 100 (output branch Y of block 116) the system 1 ? available to receive requests for prediction of malfunction of the entity? monitored 4 executed through a user device 2 (block 118).

The prediction procedure 200 (to which the flowchart of Figure 4 refers), initially, envisages acquiring a plurality of of operational data DE relating to the entity? monitored 4 (block 201). In particular, the operational data DE are acquired through the sensors 3 and/or the user device 2.

Advantageously, the operational data DE include values relating to characteristic parameters of the entity? monitored independent of time and/or relative to characteristic parameters of the entity? monitored time-dependent in a similar way as described above with respect to entity data? db.

The acquired operational data DE is then analyzed and separated into main data, or core data, and secondary data, or non-core data (block 202). Preferably, the master data comprises a subset of operational data DE corresponding to training data DT used during a training method described hereinafter. Otherwise, the secondary data includes a subset of the DE operating data that does not match the DT training data.

For example, the data processing module 12 comprises a unit? discrimination device 121 configured to analyze and separate the operational data DE into main data and secondary data.

The secondary data is processed to generate perturbation data (block 203). For example, the data processing module 12 comprises a unit? of perturbation 122 configured to receive secondary data as input and generate corresponding perturbation data as output.

The main data and the perturbation data are fed into the artificial intelligence algorithm executed by the artificial intelligence module. Does the AI algorithm provide a PM malfunction prediction of the entity? monitored 4 (block 204).

Preferably, the malfunction prediction comprises at least one probability that the entity?, or one of its components, suffers a malfunction within a future time interval starting from the acquisition of the DE operational data of the entity? monitored 4

In a preferred embodiment, the artificial intelligence algorithm is configured to provide a sequence of PM1-n malfunction predictions referred to consecutive time intervals. For example, generate a probability? of malfunction associated for two or more? selected time intervals among 6 months, 12 months, 24 months, 36 months, 48 months, 60 months starting from the acquisition of the DE operational data of the entity? monitored 4.

Furthermore, the artificial intelligence algorithm generates an indication of one or more characteristic parameters of the entity? monitored 4 results pi? relevant in the calculation of the probability? of malfunctions for each time interval considered.

? however it is clear that the examples reported above must not be interpreted in a limiting sense and the invention thus? conceived ? susceptible to numerous modifications and variations.

For example, in alternative embodiments (not shown) the operational data relating to the entity? to be monitored are provided to system 1 only through one or more? user devices 2 connected to the interface module 11 of the system 1. In this case, the measurements supplied by the sensors 3 are received by the user devices or, alternatively, are no sensors 3 provided for carrying out an automatic measurement of the parameters of the entity? monitored 4, but the parameters are determined on the basis of indirect measurements/observations.

Nothing prevents the interface module 11, the data processing module 12 and/or the artificial intelligence module 13 from being made as a single module.

Sar? clear to the person skilled in the art that the interface module 11, the data processing module 12 and/or the artificial intelligence module 13 can be implemented through both dedicated hardware resources and shared hardware resources. Again, one or more modules can be implemented, at least in part, using virtual machines.

Although the preferred machine learning algorithm is of the Random Forest type, nothing prevents you from configuring the artificial intelligence module to implement a different algorithm. In particular, in alternative embodiments (not shown), the artificial intelligence module is configured to run an XGBoost, Logistic Regression or Neural Network (NN) type algorithm.

How will it be? evident to the technician of the sector, one or more? steps of the procedures 100 and 200 described above can be performed in parallel with each other or with an order different from the one presented above. Similarly, one or more optional steps can be added or removed from one or more? of the procedures described above.

For example, in alternative embodiments (not shown), data smoothing and data separation into training set and validation set are performed in reverse order.

In one embodiment (not shown), ? omitted the feature reduction step in the training procedure. In addition or alternatively, in the forecasting procedure, the conversion of secondary data into perturbation data can be omitted.

Again, nothing prevents you from omitting the validation steps and/or the hyperparameter adjustment steps.

In an alternative embodiment (not shown), ? expected to redefine a plurality? of training data subsets with which to sequentially train the AI algorithm in order to determine a definitive training data set. In detail, after completing the AI training by means of the first subset of training data, the AI algorithm ? performed to process the data of? training set DT in order to determine the entities? training with a probability? of malfunction equal to or lower than a first threshold? for example, 20%. Subsequently, ? expected to select a new subset of entities? training, consisting of a sampling of entity? included among the entities training with a probability? of malfunction equal to or lower than the first threshold and the entity? training device malfunctioning. The artificial intelligence algorithm AI ? trained with the new subset of training data and then ? performed to process the data of? training set DT in order to determine the entity? training with a probability? of malfunction equal to or lower than a second threshold ? for example, 30%. The steps just described are iterated each time to identify entity? training with a probability? equal to or lower than a threshold higher than the previous one ? for example, three other iterations with thresholds equal to 40%, 50% and 60%, respectively ? until the number of entities? of identified training has a significant value? for example, a threshold beyond which the system degrades strongly such as over 70% in the example of the iterations considered above. Are the results of the previous steps used to determine the probability distribution? of malfunction of the entities? training data set DT and determine the final training data subset in the same way as described above.

How will it be? evident in the case of a highly unbalanced training data set, i.e. comprising a group of entities? of training of the set of relaxation data that have suffered a malfunction much less than a group of entities? data of the set of training data that have not suffered a malfunction ? for example, in which the entities? of training of the first group are 5% or less of the entities? of training of the second group - the weight of the first group of entities? of training ? on the probability distribution? malfunction of the training data set ? negligible.

Consequently, the step of determining the probability distribution? of malfunction of the entities? of training of? training set DT (block 112 of procedure 100) can? be implemented by determining the probability distribution? of malfunction of the entities? training of only the second group of entities? ? i.e. that they have not suffered a malfunction. Consequently, the step of determining a probability distribution? of malfunction of the entities? of training of the training data DT (block 113) provides for the selection of a third subset of entities? of training SDT3 of the second group? that is, all entities? that have not suffered a malfunction ? in such a way that the third subset of entities? training present a distribution of probabilities? malfunction equal to the distribution of probability? of malfunction of the entities? data of DT training data. Finally, the second subset of entities? of training SDT2 of the training data set DT ?, therefore, defined as the union of the entities? training of the first group and the third subset of entities? of training SDT3.

In a further embodiment (not shown), ? You plan to repeat training routine 100. For example, training routine 100 can? be repeated after a predetermined period of time (such as a year) from? previous iteration of the same, acquiring new data of entity? from databases.

How will it be? Obviously, the combination of the procedures 100 and 200 presented above form an overall malfunction prediction method of an entity? monitored according to the present invention.

Naturally, all the details can be replaced by other technically equivalent elements.

Also, although reference was made to an entity in the description? monitored corresponding to an industrial plant, sar? evident to a person skilled in the art that the method and system described above can be adapted to provide predictions relating to entity? different without requiring substantial modifications to the method and system according to the present invention.

For example, in one embodiment the method and the system are configured to provide failure predictions of a computer system, such as a data center, in which case the parameters include parameters related to installed hardware resources, parameters related to the operation of those resources hardware, parameters related to? building in which ? location of the data center, parameters related to an HVAC system that regulates the temperature of the data center, parameters related to the operation of said HVAC system, etc.

Again, the method and the system are configured to provide malfunction predictions, i.e. the failure, of an entity? legal, especially a company. In this case, system 1 ? connected to databases 5 also of a public financial nature ? for example, databases of the chambers of commerce or similar bodies of the country in which ? operational l?entity? monitored ? and private? for example, the AIDA database.

Data relating to bankrupt companies and non-bankrupt companies with characteristics similar to the entity? monitored ? for example, in terms of commercial sector and turnover. In particular, are a predetermined number of fiscal years acquired? for example the last five tax entities. Preferably, are companies excluded from the data set for which a number of official financial reports greater than or equal to a threshold number are not available? for example, equal to five.

The set of data cos? acquired are then divided into main data and secondary data as described above. The main data are then processed to standardize the set of data and ? performed a reduction of the characteristics, in a similar way as described above.

The secondary data is processed to generate perturbations to be fed into the artificial intelligence algorithm. In this case, the secondary data are processed according to a tree taxonomy based on the SHELL model (Software, Hardware, Environment, Liveware People and Liveware Environment) as described in Cantamessa, M., Gatteschi, V., Perboli, G., Rosano , M., 2018: ?Startups? Roads to Failure?, Sustainability 10 (7), 2346. In particular, the result of this elaboration ? processed by means of a set of rules that map the effect of the results of processing on the main data and create a matrix of impact of output (risk impact matrix) that ? provided as input to the artificial intelligence algorithm.

The set of entity data? of training derived from the main data is highly biased, in particular the number of companies that have not gone bankrupt ? substantially greater than the number of failed companies ? for example, the number of companies that did not go bankrupt ? greater than two orders of magnitude compared to the number of companies gone bankrupt. The set of data ? sampled as described above. In one embodiment, ? planned to sample only the companies not failed and keep the totality? of failed companies or, alternatively, ? expected to define a relationship between non-bankrupt companies and failed companies between 5 and 1, plus? preferably between 3 and 1.

In a similar way to what was described above, the artificial intelligence algorithm ? trained with the training data subset and with the perturbation data, validated and possibly the hyperparameters are adjusted.

After that, the entire set of training data is processed by the artificial intelligence algorithm in a similar way to what is described above. At the end of the processing by the artificial intelligence algorithm, there is a probability that of bankruptcy for each of the companies in the training set for each time interval considered, for example for each of the five years.

As described above, the probabilities? of failure are combined to obtain a probability distribution? of failure for each time interval considered. Consequentially, ? generated a new subset of training data including training data associated with a new sample of companies in the training set selected to have a probability distribution? of failure during each interval of time equal to the corresponding distribution of probability? failure rate for each time interval of the companies in the training set.

The artificial intelligence algorithm ? then trained with the new validated training data subset and eventually the hyperparameters are adjusted, completing the system training procedure.

During the prediction procedure, through the user device ? Is it possible to request a prediction from the system on the probability? of bankruptcy of a monitored company. In particular, ? expected to receive incoming corporate data and, preferably, a risk matrix related to the company directly through the user device. Furthermore, are the sensors omitted or replaced by a connection to a computer system capable of providing financial and/or productivity information? of the monitored company? for example, a real-time performance of a stock market price of the company, its client and/or affiliated companies, and/or a stock market performance of goods manufactured or traded by the monitored company.

In a similar way to what was described above, the input data is divided into main data and secondary data, with the secondary data being processed to generate corresponding perturbations which are input to the artificial intelligence algorithm together with the main data.

The main intelligence algorithm therefore provides an output prediction of failure associated with the monitored company referring to a plurality of companies. of future time periods. In one embodiment of the present invention, five predictions of failure are provided, which include a probability of failure. of bankruptcy of the company at 12 months, 24 months, 36 months, 48 months and 60 months, from the request for prediction. Preferably, one or more predictions are also provided together with the failure prediction. performance indicators related to national and international regulations ? for example Altman's z-score and/or Italian Business Crisis Alert Indices (defined in the Business Crisis and Insolvency Code - Legislative Decree 14/2019).

In a preferred embodiment, the failure predictions and input data are further processed so as to identify the parameters that have the greatest impact on the probability of failure. of bankruptcy of the monitored company.

Finally, the prediction of probability? and the identified critical parameters are combined into an output report that ? provided to the user device that made the request.

Claims (10)

1. Method (100, 200) to predict a malfunction of an entity? (4), said entity? (4) being characterized by a plurality? of parameters, the method (100, 200) providing for: predict (200) a malfunction of the entity? (4) by means of an artificial intelligence system (1) which is provided with data relating to the parameters of the entity? (4), in which the artificial intelligence system (1) ? trained using training data characterized in that training data is selected according to the following steps: to. acquire (101-106) a set of training data, wherein the set of training data comprises a plurality of? of entity data groups, each entity data group? including values referring to characteristic parameters of a respective entity? of training, b. select (107) entity data? included in entity data groups? of a first subset of entities? data set training data, c. train (109) the artificial intelligence system using entity data? of the first subset of entities? of selected training, d. determine (112) a probability distribution? of malfunction of the entities? of training of the training data set, running the artificial intelligence system (1) trained with the data of entity? of the first subset of entities? of training, And. determine (113) the training data by selecting data of entity? of entity? training belonging to a second subset of entity? of training of the set of training data in such a way that the second subset of entities? training present a distribution of probabilities? of malfunction equal to that calculated in point d. The method (100; 200) according to claim 1, further comprising the step of: - determine a first group of entities? of training of the set of data of slack that have suffered a malfunction and a second group of entity? data set of the training data set that have not suffered a malfunction, where the second group ? much greater than the second group: the step of determining (112) a distribution of probabilities? of malfunction of the entities? of the training dataset includes: - determine the distribution of probabilities? of malfunction of the entities? of training of the second group, e in which the step of determining (113) the training data by selecting data of entity? of a second subset of entities? of the training dataset includes: - select a third subset of entities? training of the second group in such a way that the third subset of entities? training present a distribution of probabilities? of malfunction equal to that calculated in points d, e - define the second subset of entities? of training of the set of training data as a union of the entities? of training of the first group and of the third subset. The method (100; 200) according to claim 1 or 2, wherein the step of determining (112) a probability distribution? of malfunction of the entities? of training of the training set includes: the. define a plurality? of probability threshold values; ii. determine at least one of a number and a list of entities? of training of? set of data of training whose probability? of malfunction? less than or equal to a selected threshold value, running the AI system, iii. to select a new threshold value of probability? of said plurality? of probability threshold values?, iv. iterate steps i - iii, until all probability threshold values have been selected, e v. calculate the probability distribution? of malfunction of the entities? training set based on the results provided at each iteration of step ii. 4. Method (100; 200) according to claim 3, further provides the steps of: ii_a. select entity data? of an intermediate subset of entities? training, said intermediate subset of entity? training including entity? included in the list of entities? of training determined in step ii and entity? of training set that have suffered a malfunction; ii_b. train the artificial intelligence system using entity data? of the intermediate subset of entities?, before performing steps iii - v. The method (100; 200) according to claim 1 or 2, wherein the step of determining (112) a probability distribution? of malfunction of the entities? of training of the training set includes: - calculate a probability? of malfunction for each entity? of training of the training data set, running the artificial intelligence system, e - combine the odds? malfunction calculated to determine the distribution of probabilities? of malfunction of the entities? of training data set. 6. Method (100; 200) according to any one of the preceding claims, wherein the step of determining (112) a probability distribution? of malfunction provides for: - calculate a plurality? of probability distributions? of malfunction of the entities? of training of? set of training, where each distribution of the probabilities? of malfunction of the entities? of training of? training set ? calculated by means of groups of data of the entities? of training referred to the same time interval, e wherein the step of determining (113) the training data provides for: - select entity data? of a second subset of entities? of training of the set of training data in such a way that the second subset of entities? of training present a plurality? of probability distributions? of malfunction equal to the plurality? of probability distributions? of malfunction calculated for the entities? of training data set. The method (100; 200) according to any one of the preceding claims, wherein the step of selecting (107) entity data? of a first subset of entities? of the training dataset includes: - determine a first group of entities? data set training data that has malfunctioned; - randomly sample a second group of entities? training group of the set of training data that have not suffered a malfunction, where the second group includes a number of entities? training between 1 and 5 times a number of entities? of training of the first group, and - compose the first subset of entities? of training of the set of training data as a union of the entities? of training of the first group and of the second group. The method (100; 200) according to any one of the preceding claims, wherein the step of predicting (200) a malfunction of the entity? (4) by means of an artificial intelligence system (1) plans to use an automatic learning algorithm, selected from: -Random Forest, - XGBoost; - Logistic Regression, e - a neural network. The method (100;200) according to any one of the preceding claims, wherein the step of predicting (200) a malfunction of the entity? (4) by means of an artificial intelligence system (1) plans to: - calculate a plurality? of probability of malfunction of? Entity? (4), each probability? of malfunction of? Entity? (4) referring to a respective future time interval. 10. System (1) for predicting a malfunction of an entity? (4), said entity? (4) being characterized by a plurality? of parameters, the system (1) including: - an interface module (11) configured to receive data relating to said entity? (4) and provide a malfunction prediction result; - a data processing module (12) configured to acquire and process data relating to entities? of training and the data relating to this entity? (4) supplied to the interface module (11), e - an artificial intelligence module (13) configured to execute an artificial intelligence algorithm which uses the data processed by the data processing module (12) to predict a probability? of malfunction of? Entity? (4), wherein the modules (11, 12, 13) of the system (1) are configured to implement the method according to any of the preceding claims.